What Changed in Claude Fable 5.1: Coding Goes 42% to 55.8%, Forced Tool Calls Return a 400

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 is open to every Claude API customer under the model ID claude-fable-5-1. Mythos 5.1 is limited to participants in Project Glasswing, a program for United States organizations. The Hacker News announcement thread drew 648 points and 604 comments.

Here is what changed from Fable 5, released in June, in order. There are four groups. Capability rose on all nine published benchmarks, five things are new in the API, the cache read price dropped to a quarter, and three behaviors break with a 400 when you migrate. The 1M token context window, the 128k max output, and the tokenizer are unchanged.

Capability: all nine benchmarks moved up

The platform docs name six areas where the gains concentrate.

• Agentic coding over long sessions. Multi-file features, large refactors and migrations, debugging, and code review across sessions that run for hours.
• Documents, spreadsheets, and slides. Taking an analysis from a first question to a finished document, a live-formula spreadsheet, or a deck built from a blank page.
• Research and search. Higher accuracy on multistep web research that follows up on what it finds.
• Vision. Reading dense charts, filings, and tables nested inside PDFs, including with crop-and-zoom tools on charts.
• Long-context work. Connecting details across the full 1M token window.
• Computer use. Driving a browser and desktop apps more reliably, and recovering from failed steps.

Multilingual performance is on par with Fable 5. The docs attach one condition to all of it: the gap widens at higher effort levels. A pipeline pinned to low effort to save money may not see much of this release.

For numbers, the announcement puts Fable 5.1 next to its predecessor and to Claude Opus 5, the model the docs still recommend as the starting point. Fable 5.1 leads both on all nine. The two biggest jumps are in work that runs for a long stretch: Terminal-Bench-Science 0.1, which carries a research task to completion in a terminal, more than doubles from 24.7% to 52.6%, and AutomationBench, covering business workflow automation, goes from 17.1% to 31.4%. Agentic coding on Terminal-Bench 4.0 moves from 42.0% to 55.8%.

Mythos 5.1, running under looser cyber safeguards, hits 60.9% on that same Terminal-Bench 4.0. The two are the same underlying model; the gap is the set of tasks where the safety classifier steps in.

Speed has a number attached too. Citing external partner testing, Anthropic says Fable 5.1 ran roughly twice as fast as Opus 5 and used half as many tokens. The baseline there is Opus 5, not Fable 5. In Claude Code, cybersecurity-related false positives dropped roughly 60%, meaning that much more security work gets through instead of being refused. All of these benchmarks and quotes are Anthropic's own selection, though, and a day after launch there is no independent reproduction yet.

Five things that are new in the API

The docs group five items as additive. Three are API features behind beta headers, one is pricing, one is how the output is marked.

First, you can change the thinking effort mid-conversation. Raising or lowering effort partway through a conversation no longer invalidates the prompt cache you have built up. Turn it up for a hard step, down for routine ones, inside a single session. It needs the mid-conversation-output-config-2026-07-01 beta header, and Fable 5.1, Mythos 5.1, and Opus 5 support it.

Second, system messages can be scoped to a single turn. Set clear_at: "next_user_message" on a role: "system" message and its text carries system-prompt authority for the current turn, then stops rendering once a later user message exists. The message stays in messages and you keep sending it back verbatim, so nothing earlier in the conversation changes, and a cleared message costs no input tokens. The docs pitch it as the replacement for per-turn reminders you inject into history and delete on the next request, like "check your inbox before running more code." The beta header is mid-conversation-system-clear-at-2026-08-21.

Third, you can read the progress updates between tool calls. The model writes a short note on what it found and what it will do next right before each tool call. Under the default thinking.display: "omitted" those blocks come back empty, so a long agentic turn can look silent to your users. Setting display: "updates" returns just those status lines while reasoning stays hidden. The beta header is thinking-display-updates-2026-08-18.

Fourth, cache reads got cheaper. That has its own section below.

Fifth, outputs carry provenance marks. Text carries the statistical watermark applied to models released after August 2, and images and video the model produces, through the code execution tool for example, carry signed C2PA Content Credentials when retrieved through the Files API. The docs state the watermark adds no tokens or hidden characters and carries no information about you or your organization. The detection API that reads it opened in private preview with this release, but access is restricted to regulators, law enforcement, journalists, fact-checkers, independent researchers, educational institutions, and EU civil society organizations, so ordinary developers still cannot check their own output.

The first three are for code that builds the messages array itself. They are not settings a Claude Code or claude.ai user can switch on.

Only the cache read price went down

Several outlets led with "up to 45% cheaper." Open the pricing table in the platform docs and base input sits at $10 per million tokens with output at $50, the same numbers as Fable 5. Exactly one line item moved: cache reads.

Start with what a cache read is. A long conversation or agent session resends its entire opening every time: the system prompt, the tool definitions, and the transcript so far. Prompt caching stores that opening on the server and reuses it on the next request. The rate charged on those reused tokens is the cache read price. On other Claude models it is 0.1x base input; on Fable 5.1 and Mythos 5.1 it is 0.025x. With base input at $10, a million cached tokens that cost $1 now costs $0.25.

The announcement quotes two savings figures: roughly 25% against Fable 5 on typical work, and up to roughly 45% on complex coding and agent workloads. The arithmetic explains the gap. Savings scale directly with the share of your input tokens that come from cache, so a long agent session that rereads the same opening dozens of times approaches 45%, while a one-shot call that sends a fresh prompt every time saves nothing. Cache writes are unchanged at $12.50 for five minutes and $20 for an hour, the minimum cacheable prompt is still 512 tokens, and batch processing runs $5 input and $25 output.

Seven behaviors that change without a code edit

These are the items the docs group under "Changed from Claude Fable 5." Each ships with a prompting fix in the docs.

• Parallel tool calling is more variable. Where Fable 5 batched several independent reads into one call, Fable 5.1 sometimes issues one per turn. Answer quality holds, but round trips, tokens, and wall-clock time go up. It shows up most in hand-built coding agents, bash-and-editor harnesses, and computer use. The fix is a one-line batching instruction in the prompt.
• Whole-file rewrites for small edits. The result is usually the same, but the rewrite costs more output tokens and time, and output runs $50 per million.
• Fewer progress updates during long tool runs. More so at higher effort. The display: "updates" option above is the counter-move.
• Answers from memory more often at low effort. Raise effort for turns that need fresh information.
• Denser prose. Longer sentences, fewer paragraph breaks.
• Less formatting in chat. Bold, headers, and lists appear less often, so anti-formatting rules written for older models can suppress structure the content needs.
• Unmarked quotations in summaries. If you pipe summaries straight to an external audience, quotation marking now needs its own check.

This list is assembled for migration, so it collects only what got worse. It shipped in the same release as the benchmark gains above, and which side you feel first depends on your workload: heavy iterative edits to single files surface the extra output tokens, while long research and automation runs surface the score increases.

What returns a 400 when you migrate

This is not a swap-the-model-ID update. The docs name three behaviors that break coming from Fable 5.

First, forcing the model to call a tool is now rejected. Sending tool_choice as {"type":"any"} or {"type":"tool","name":"..."} returns a 400.

tool_choice: type "tool" and "any" are not supported for this model.

The token counting endpoint applies the same validation. The default {"type":"auto"} and {"type":"none"} still work. The reason is in the docs: this model always has reasoning on, and forcing a tool call makes it skip that reasoning and write the working-out into the tool arguments, which degrades argument quality. If you were forcing a tool to pin output to a JSON schema, move to strict tool use with strict: true or to structured outputs. If you were forcing it so the tool actually gets called, state in the prompt when the tool applies.

Second, earlier models cannot read thinking blocks that Fable 5.1 produced. Preservation runs one direction only. A conversation that moves onto Fable 5.1 from Opus 5 or Fable 5 keeps its reasoning; moving from Fable 5.1 down to another model loses it for the turns that run there. If a router or fallback swaps models mid-conversation, the API silently discards blocks it cannot read. Dropped blocks are not billed and do not count toward input_tokens, but without the thinking-binding-controls-2026-08-01 beta header the response never says the drop happened.

Third, editing an earlier turn invalidates every thinking block after it. Change the system prompt, the tools array, or any prior message, and the next request is rejected with a 400 carrying The block is bound to a different conversation. The check has a grace period: it is enforced only on accounts created on or after August 31, 2026, while older accounts merely log mismatches unless the request sets thinking.block_binding.prefix_mismatch_behavior. Mythos 5.1 does not run the check at all.

The things that already returned a 400 on Fable 5 still do: thinking: {"type":"enabled"} and {"type":"disabled"} both error, as do assistant response prefills and non-default temperature, top_p, or top_k. All three breaking changes apply only to code that builds the messages array itself. Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK keep the prefix stable on their own. If you only use the official tooling, there is nothing to change here; this section is for developers calling the API directly.

Can you use it today

Fable 5.1 is available worldwide, Korea included. Neither the Claude API, claude.ai, AWS, Google Cloud, nor Microsoft Foundry lists a regional restriction. Mythos 5.1 is open only inside Project Glasswing, which admits United States organizations only.

The Fable 5 promotional credits that expired on July 19 do not carry over; the support docs state Fable 5.1 was never in scope for them. Retention terms need checking too: both models retain data for 30 days and cannot run under a zero data retention configuration unless Anthropic expressly authorizes it. For teams whose internal policy requires no retention, that line decides adoption before anything else does.

One more thing: the docs do not recommend Fable 5.1 as a default. They say most workloads should start on Claude Opus 5, and to move up to Fable 5.1 only when Opus 5 at xhigh or max effort still falls short on your own evaluations.

If you already call Fable 5 from the API, grep the codebase for tool_choice and clear out any and tool before you touch the model ID, or your first request comes back a 400. If you are not on the Fable line yet, run your evals on Opus 5 first and reach for Fable 5.1 only where those fall short. Either way, do not leave effort at its high default without retuning it, which is step three of the official migration guide.